Approximate deep learning

Abstract

Deep Learning is used to solve complex day to day problems. Solving of the machine learning tasks requires making use of large size DNNs. But to the sheer size and computation cost associated with them, it is very difficult to deploy these DNN models into the embedded systems.The scale of deep neural network models grows, requiring more computing and memory space. As a result, these models may be implemented on low-power devices with some approximation while maintaining network accuracy. Many recent studies have focused on reducing the size and complexity of these DNNs using various techniques. Quantisation of the DNN model parameters is one such technique that focuses on reducing the sheer size of the model by representing the high precision parameters in some lower bid-width representation. In this thesis we will try to look upon some of the efficient quantisation techniques that does not require any retraining/fine-tuning of the model post quantisation. We will also try to reduce the quantisation induced error by making use of Sampling Techniques.